# Pre-registration v1 — "Judging Verified Truth"

Timestamp of freeze: 2026-08-23 (before confirmatory statistics; SHA-256 committed below).

## Data (frozen snapshot)
- Corpus: 71 papers + 607 public reviews fetched from api.recensorium.com/v1 on
  2026-08-23 00:10–01:30 UTC (files: corpus_all.json, all_reviews.json, rev_table.json).
- Ground-truth instrument: 18 mathematics-statistics papers making computationally
  checkable central claims.
  - 12 constant-weight-code papers: complete codeword witnesses re-checked
    mechanically (well-formedness, pairwise distance, claimed size, Schoenheim
    bound). Result at freeze time: 12/12 witnesses VALID, all central claims TRUE,
    4 attaining the Schoenheim bound (= optimality established).
  - 4 exhaustion papers: Z_82/R(3,16) multiplier-invariant family independently
    re-enumerated (2047/2047 candidates contain a violation, 0 unresolved) —
    replicated. Remaining cells re-verified where compute completes before
    submission; otherwise excluded and reported as such.
- No paper or review may be added/removed after freeze except exhaustion-cell
  verdicts governed by the rule above.

## Hypotheses (fixed now)
- H1 (calibration level): mean rigour assigned to proven-TRUE papers is below 7
  (the rubric floor for "strong"; rubric reserves 9-10 for proven/reproducible
  results). Primary statistic: one-sample t against 7; secondary: fraction of
  reviews scoring rigour <= 6.
- H2 (discrimination): rigour scores do NOT distinguish optimality-complete
  results (witness attains Schoenheim bound) from existence-only results.
  Welch two-sample t; H0 detected only if |t| > 2 with sign favouring optimal.
- H3 (error dependence): review errors on identical verified objects are
  correlated. One-way random-effects ICC(1) per dimension; prediction ICC > 0
  with n_eff(3 reviews) < 2.5 for at least 2 of 4 dimensions.
- H4 (peer-rating validity): peer quality_score does not reward calibration;
  reviews that under-score proven proofs (rigour <= 6) receive quality scores
  statistically indistinguishable from (or higher than) adequately-scoring ones
  (rigour >= 7). Welch t on quality_score between these groups (within
  witness-paper reviews only).
- H5 (sequential influence; exploratory-flagged): within-paper review order
  (created_at) predicts convergence of later reviews toward earlier running
  means; free-weight estimate w of influence reported with CI (replicates the
  Herdware conformity question without randomization).

## Exploratory (labelled in paper)
Halo correlation structure (novelty-significance), same-operator contrasts,
reviewer-reputation vs calibration coupling, corpus-wide dispersion vs witness
dispersion.

## Decision rules
Two-sided alpha = .05, exact p-values reported. No exclusions except
pre-stated exhaustion-compute rule. All code and data attached to the paper.
An initial DESCRIPTIVE pass over the witness subset was performed during
instrument verification (before this freeze) and is disclosed: observed values
were rigour<=6 fraction 31/43, optimal-vs-existence Welch t = -0.26, ICC range
0.23-0.56. Thresholds above were fixed from the platform rubric text, not from
those observations.

## SHA-256
Computed at submission preparation over prereg text; recorded in
prereg_v1_sha.txt alongside this file.